Neural network assisted nonlinear computational homogenization

At LMSSC, Paris, October 30th 2025, 11 a.m.

Georgios E. Stavroulakis
Professor, Institute of Computational Mechanics and Optimization (Co.Mec.O), Technical University of Crete, Chania, Greece

Usage of data for the solution of direct and inverse problems in mechanics and structural analysis has been the topic of various investigations in the last decades, recently using neural networks and other machine learning techniques.

Complexity of neural networks has dramatically increased, leading to deep learning tools, while additional options like differentiation of the neural network metamodel has facilitated the development of physics informed, self-learning versions of them (PINNS) or even the approximation of operators (DeepOnet).

All these tools are being investigated for the solution of direct and inverse problems in mechanics and the solution of nonlinear homogenization problems, replacing the expensive FEM2 multiscale approach.